← back
Demand-Driven Context: A Methodology for Coherent Knowledge Bases Through Agent Failure
Takeaway
Stop monolithic retrieval layers; decompose institutional knowledge into demand-driven context blocks discovered from real agent failures.
Summary
- IKEA staff engineer cites McKinsey: 88% of enterprises use AI but only 6% see value creation — Jira epics aren't moving despite agentic hype.
- Maps task types: green (general knowledge LLM has), orange (teachable skills), red (institutional/tribal knowledge) — red is what blocks agentic delivery.
- Critiques the rush to build 10-20 MCP servers and RAG layers as solving symptoms; outputs are non-deterministic and rarely evaluated.
- Enterprise knowledge is a monolith: 20% outdated, 20% unreliable, 10% duplicated, 40% tribal — must be decomposed into context blocks like a microservices migration.
- Demand-driven context methodology builds the knowledge base from observed agent failures rather than trying to document everything upfront.
agentsknowledge-basemcp
Original description
Enterprise teams spend a lot of time trying to guess what AI agents need to know. This workshop flips that around. Instead of curating context top-down, Raj Navakoti shows how to build a demand-driven context base by giving agents real problems, watching where they fail, and using those failures to reveal exactly what knowledge is missing. Using practical exercises and real examples from IKEA Digital, the session walks through how to grow a knowledge base problem by problem, structure it in Markdown, and use agents with different roles and reasoning boundaries against the same shared context. If you're building enterprise AI systems and want a more grounded way to create useful context, this is a strong practical framework. Speaker info: - https://www.linkedin.com/in/raj-navakoti-529880b1/ Timestamps: 0:00 - Introduction and speaker background 2:47 - The situation: Analogy to the movie Memento and AI's memory constraints 3:55 - Evolution of AI: From prompt engineering to deep agents 4:33 - Enterprise AI challenge: Why productivity isn't moving 5:33 - The problem: Green (general), Orange (taught), and Red (institutional/tribal) knowledge 10:11 - The Monolith: Why institutional knowledge is often outdated or missing 11:24 - Solution introduction: Demand-driven context 13:05 - The "Pull" strategy: Learning by doing vs. pushing information 14:48 - The agent lifecycle: Problem to discovery to documentation 17:46 - Demo introduction: Using a framework for context management 19:12 - Live demo: Incident root cause analysis and context discovery 24:05 - Scaling: 14 incidents to show confidence level improvement 26:27 - Automated scale: Validating knowledge across the monolith 33:01 - Storage strategy: Why GitHub is preferred for knowledge repositories 34:47 - The Meta Model: Navigating domain relationships 36:27 - Value proposition: Knowing the unknown and managing knowledge 39:02 - Summary: The 80/20 rule and cache-based context blocks 40:15 - Workshop takeaways: Repositories and scanners 43:33 - Q&A Session: Addressing scalability, tooling, and cost